Implementation and evaluation of hardware/software performance monitor for parallel programs
Several attempts have been made to construct integral support for parallel program analysis in multiprocessor systems. The literature on the design tradeoffs encountered in such an effort is sparse as few experiments have been attempted to explore them. This is especially true for issues related to the design of the event collection mechanism. Further, many studies of program analysis tools are directed toward a specific programming style or multiprocessor architecture. This makes it difficult to compare tools and techniques. The method of making collected performance information available to the programmer often overshadows the discussion of the techniques used to collect the information. This thesis examines the trade-offs encountered in the design of a general purpose event collection tool for parallel program analysis. The discussion includes a definition of the event collector design problem and a description of the approaches taken by other efforts. An integral event collection system implemented in hardware as part of the M31 VAX multiprocessor is described and used to evaluate the most promising event collection techniques. The insight from the design and use of the M31 instrumentation system and the evaluation of collection techniques is then applied to describing the architecture of an economical hybrid collector that may be constructed as part of a shared bus multiprocessor.
- Research Organization:
- Carnegie-Mellon Univ., Pittsburgh, PA (USA)
- OSTI ID:
- 6093325
- Country of Publication:
- United States
- Language:
- English
Similar Records
Run-time support for parallel functional programming on shared memory multiprocessors
Performance evaluation of parallel algorithms and architectures in concurrent multiprocessor systems. Final report, January-September 1987